Truck Rest Stop Imputation From GPS Data: An Interpretable Activity-Based Continuous Hidden Markov Model

The increasing wealth of truck global positioning system (GPS) data has broadened the opportunities for understanding freight logistics activities and enhancing research capabilities to real-world case studies. A very important piece of information from the planning and regulatory perspective concerns the occurrence and location of rest stops. In this study, we propose a data-driven unsupervised machine learning method to impute truck stop events by using a Continuous Hidden Markov Model (CHMM). Specifically, we estimate the joint probability distribution of a mixture of multivariate Gaussian densities, whose parameters depend on the latent states of a Markov chain. Each density represents a cluster of stops that are identified not only from their spatial proximity but also from their temporal proximity as the clustering of the rest stops depends on latent states that are conditional on expected times retrieved from the observed data. In this study, we applied the proposed method to a database containing more than 71 million GPS records of Australian trucks, and we particularly aimed to identify rest stops based on a list of features related to the locations and the load of the trucks. The results showed that the CHMM could account for the location proximity for different activities of truck drivers, and they were validated against complementary data on truck loads and land use by using a stratified sampling technique. Validation results indicated that 94.1% of the rest stops were correctly identified, and highlighted the advantage of the proposed approach without any requirement of labelled data, driver logbook or complimentary survey.